Antiplex-Instruct-3B

A warm, direct, open‑world conversational AI built on Phi‑3‑mini — no corporate bot vibes, just honest chat.

⚠️ Important

This model has been completely rebuilt from the ground up. The previous version suffered from corrupted config files, fused-weight mismatches, and gibberish output. Those issues are now fully resolved.
You can load the model directly with AutoModelForCausalLM.from_pretrained — no special libraries, no hacks, no "as an AI" deflections.
Please review the model files (config.json, model.safetensors, and tokenizer files) before installation to ensure you are using the latest version. MODEL work done.

---

📋 Overview

Antiplex-Instruct-3B is a high-performance instruction-tuned language model developed by QuantaSparkLabs. Released in 2026, this model is engineered for dual-task capability, delivering accurate identity alignment, reliable SQL generation, and strong general reasoning, while remaining lightweight and efficient.

The model is fine-tuned using LoRA (PEFT) on curated datasets emphasizing identity consistency and structured reasoning, making it ideal for edge deployment and specialized assistant roles.

✨ Core Features

🎯 Task Versatility	⚡ Performance Optimized
Text Generation: SQL/NLP, creative writing, technical explanations.	LoRA Fine-tuning: Efficient parameter adaptation.
Classification: Intent detection, task routing, safety filtering.	Identity Alignment: Consistent persona across interactions.
Dual-Mode: Single model handling generation + classification.	Lightweight: ~3.8B parameters, edge-friendly VRAM footprint.

---

📊 Performance Benchmarks

🏆 Accuracy Metrics

Task	Accuracy	Confidence
Identity Verification	100%	⭐⭐⭐⭐⭐
SQL Generation	100%	⭐⭐⭐⭐⭐
General Reasoning	90%	⭐⭐⭐⭐

🔬 Reliability Assessment

21-Test Internal Validation Suite

• Passed: 16 tests (76.2%)
• Failed: 5 tests (23.8%)
• Overall Grade: B (Good)

  

📈 View Detailed Test Categories

Category	Tests	Passed	Rate
Identity Tasks	7	7	100%
SQL Generation	6	6	100%
Reasoning	5	3	60%
Classification	3	2	66.7%

Test Dataset: QuantaSparkLabs/antiplex-test-suite

---

🏗️ Model Architecture

Training Pipeline

graph TD
    A[Base Model Phi-3-mini] --> B[LoRA Fine-tuning]
    B --> C[Task-Specific Heads]
    C --> D[Text Generation Head]
    C --> E[Classification Head]
    D --> F[Generation Output]
    E --> G[Classification Output]
    H[Instruction Dataset] --> B
    I[SQL Dataset] --> B
    J[Identity Dataset] --> B

Inference Flow

User Prompt → Tokenization → Antiplex Core → Task Router 
                ↓
       [Generation/Classification] → Post-processing → Output

---

🔧 Technical Specifications

Parameter	Value
Base Model	unsloth/Phi-3-mini-4k-instruct-bnb-4bit
Fine-tuning	LoRA (PEFT)
Rank (r)	16
Alpha (α)	32
Optimizer	AdamW (β₁=0.9, β₂=0.999)
Learning Rate	2e-4
Batch Size	8
Epochs	3
Total Parameters	~3.8B

Dataset Composition

Dataset Type	Samples	Purpose
Identity Alignment	30	Consistent persona training
SQL Generation	300	Structured query training
Instruction Tuning	2,500	General capability enhancement
Classification	1,000	Intent detection training

---

💻 Quick Start

Installation

pip install transformers torch accelerate

Basic Usage (Text Generation)

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "QuantaSparkLabs/Antiplex-instruct-3B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

prompt = "Write an SQL query to fetch users created in the last 30 days."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Classification Mode

# Intent classification example
classification_prompt = """[CLASSIFY]
User Query: "I need to reset my account password"
Categories: account_issue, technical_support, billing, general_inquiry
"""

inputs = tokenizer(classification_prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=64,
    temperature=0.3,
    do_sample=False
)

detected_intent = tokenizer.decode(outputs[0], skip_special_tokens=True).split('[')[-1].split(']')[0]
print(f"Detected Intent: {detected_intent}")

Chat Interface

from transformers import pipeline

chatbot = pipeline(
    "text-generation",
    model=model_id,
    tokenizer=tokenizer,
    device=0 if torch.cuda.is_available() else -1
)

messages = [
    {"role": "system", "content": "You are Antiplex, a helpful AI assistant specialized in SQL and classification tasks."},
    {"role": "user", "content": "Classify this intent: 'Can you help me with invoice generation?' Then write a SQL query to find recent invoices."}
]

response = chatbot(messages, max_new_tokens=512, temperature=0.7)
print(response[0]['generated_text'][-1]['content'])

---

🚀 Deployment Options

Hardware Requirements

Environment	VRAM	Quantization	Speed
GPU (Optimal)	8-12 GB	FP16	⚡ Fast
GPU (Efficient)	4-6 GB	INT8	⚡ Fast
CPU	N/A	FP32	🐌 Slow
Edge Device	2-4 GB	INT4	⚡ Fast

Cloud Deployment (Docker)

FROM pytorch/pytorch:2.0.1-cuda11.7-cudnn8-runtime

WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

COPY . .
EXPOSE 8000

CMD ["python", "app.py"]

---

📁 Repository Structure

Antiplex-Instruct-3B/
├── README.md
├── model.safetensors
├── config.json
├── tokenizer.json
├── tokenizer_config.json
├── generation_config.json
├── special_tokens_map.json
├── quantasparklogo.png
├── examples/
│   ├── classification_demo.py
│   ├── sql_generation_demo.py
│   └── chat_interface.py
└── evaluation/
    └── test_results.json

---

⚠️ Limitations & Safety

Known Limitations

• Domain Specificity: Not trained for medical/legal/safety-critical domains
• Bias Inheritance: May reflect biases in training data
• Context Window: Limited to 4K tokens
• Multilingual: Primarily English-focused

Safety Guidelines

# Recommended safety wrapper
def safety_check(text):
    blocked_terms = ["harmful", "dangerous", "illegal", "exploit"]
    if any(term in text.lower() for term in blocked_terms):
        return "Content filtered for safety reasons."
    return text

---

🔄 Version History

Version	Date	Changes
v1.0.0	2026-01-1	Initial release
v1.1.0	2026-01-10	Enhanced classification head
v1.2.0	2026-01-25	SQL generation improvements

---

📄 License & Citation

License: Apache 2.0

Citation:

@misc{antiplex2026,
  title={Antiplex-Instruct-3B: A Dual-Task Instruction-Tuned Language Model},
  author={QuantaSparkLabs},
  year={2026},
  url={https://huggingface.co/QuantaSparkLabs/Antiplex-instruct-3B}
}

---

👥 Credits & Acknowledgments

• Base Model: Microsoft Phi-3 Mini team
• Fine-tuning Framework: Unsloth for efficient LoRA training
• Evaluation: Internal QuantaSparkLabs team
• Testing: Community contributors

---

🤝 Contributing & Support

Reporting Issues

Please open an issue on our repository with:

1. Model version
2. Reproduction steps
3. Expected vs actual behavior

---

Built with ❤️ by QuantaSparkLabs
_{Model ID: Antiplex-Instruct-3B • Parameters: ~3.8B • Release: 2026}

Someone gimmi a cup of coffe!☕